Firstly, You have to determine which versions you want to use for each library and NVIDA driver.

1. Determine the specific TensorFlow version.

2. Figure out the compatibility with dependent libraries.
   Refer to https://www.tensorflow.org/install/source_windows#gpu

3. Figure out the compatibility with dependent NVIDA driver.
   Refer to https://docs.nvidia.com/deploy/cuda-compatibility/index.html#binary-compatibility

$ nvidia-smi
Thu Mar 04 13:00:19 2021       
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 442.62       Driver Version: 442.62       CUDA Version: 10.2     |
|-------------------------------+----------------------+----------------------+
| GPU  Name            TCC/WDDM | Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|===============================+======================+======================|
|   0  GeForce GTX 1060   WDDM  | 00000000:01:00.0  On |                  N/A |
| N/A   49C    P8     4W /  N/A |    831MiB /  6144MiB |      0%      Default |
+-------------------------------+----------------------+----------------------+

This Driver is included within CUDA toolkit of the next section.
Before installing the driver you have to check out the above “Pre-check up” section.
If not have to do it, you’d better you don’t install the driver.

You can install CUDA library with conda.
Furthermore, you can do several cuda versions independently.
It’s more convenient and has some advantages, you’d better use conda if there are not any other matters.

The way of using conda, you can see the next sub-pages.

If NVIDA driver has been already installed as what is satisfied, unselect the driver into the custom install step.